Data Normalization, Denormalization,

and the Forces of Darkness

a white paper
by Melissa Hollingsworth
melissa@fastanimals.com
http://www.fastanimals.com/melissa/

v. 0.4

This paper is in beta. Feedback is welcome.

 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p1

Introduction
According to one SQL blogger, “normalization” and “denormalization” are the most common
search terms bringing people to his site. Typing “normali” into Google brings up suggestions
like “normalization” and “normalizing data” high on the list. If you actually search for
“normalization,” your top search results include Wikipedia, overviews, tutorials, and basics. All
the bases are covered!
Many of these overviews just skim the surface, never explaining why anyone would bother
doing this. Some use examples which illustrate one principle while violating others, leading to
confusion. Many use precisely the same examples for the higher forms, reworded slightly from
each other, which might lead a more cynical person to wonder how well the writers grasp what
they describe. If we can’t create our own examples, do we really understand the principle?
I do see a need to distribute basic information more widely. If it were common knowledge,
we wouldn’t see so many people searching for it. We also wouldn’t see so many people
complaining that it’s hard to understand.
This paper covers the basics in a manner which I hope is both accurate and easy to
understand, and also slithers into intermediate territory. We need at least an intermediate level of
knowledge before we can decide how far to normalize, the possible efficiency tradeoffs, if
denormalization is needed, and (if it is) what denormalization methods to use.

What Is Normalization?
“Normalization” just means making something more normal, which usually means bringing it
closer to conformity with a given standard.
Database normalization doesn’t mean that you have have weird data, although you might. It’s
the name for an approach for reducing redundant data in databases. If the same data is stored in
more than one place, keeping it synchronized is a pain. Furthermore, if you can’t trust your
synchronization process absolutely, you can’t trust the data you retrieve.
You may have heard that normalization is the enemy of search efficiency. This is kind of
truish, but not really true. At the lower levels, imposing a bit of order on your chaotic heap of
data will probably improve efficiency. At the higher levels you may see a performance
slowdown, but it probably won’t be as bad as you fear and there are ways to address any
problems you do find.
Let’s get to work on a data set.
I work for a company which sells widgets. We must track widget inventory, prices, order
numbers, blah blah blah . . . . Lord, I’m falling asleep just writing about it.
You know what? You look trustworthy. I’m going to tell you the sinister truth. I’ve been
retained by an organization called the Council of Light, which stands alone against the forces of
evil and darkness which daily threaten to engulf our innocuous, daylit world.
 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p2

The Council of Light wants me to put their data into a SQL database so that they can retrieve
it more easily. They have lists of monsters, types of weapons, historical records of monster
outbreaks, a registry of currently available monster fighters, and a great deal more.
Unfortunately, it’s all stored in a combination of dusty tomes and piles of index cards. I must
design a database which models this information accurately and usefully.
Let’s dive in.
A table of active monster fighters seems fairly obvious—full name, date of birth, formal
training, date of joining the Council, fighting skills, and whatever else we come up with.

DATE
FIRST DATE OF
LAST NAME TRAINING JOINED FIGHTING SKILLS
NAME BIRTH
COUNCIL
blades, distance
Knights Templar weapons, firearms,
Buttkicker Brenda 19850123 20050615
Special Ops martial arts,
munitions, stakes
Cutthroat Cathy 19760721 Order of the Stake 20020513 blades, stakes
blades, distance
Killer Karl 19840812 Demon Dojo 20040914 weapons, martial
arts
blades, distance
Martyr Mel 19790304 Demon Dojo 20000912 weapons, martial
arts
Monsterbasher Marvin 19820118 Order of the Stake 20071120 blades, stakes
blades, firearms,
Vampkiller Valerie 19810329 School for Slayers 20030502
stakes
And 50
others...

And I’d better have a table of those who provide support from the sidelines—students of the
arcane, rogue fighters who sometimes partner with the Council, etc. Two tables so far.
 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p3

DATE OF DATE JOINED

LAST NAME FIRST NAME EXPERTISE
BIRTH COUNCIL
antique weapons,
American folklore,
Clueful Clarence 19640723 19950312
American history,
blades, martial arts
world folklore, world
Elder Ellen 19440105 19680345
history
demons, fey,
Learned Lucy 19510821 19730103 lycanthropes,
undead
Rogue Rodney 19840323 firearms, munitions
lycanthropes,
Sage Sally 19710612 19980226 Romanian folklore,
Romanian history
And 200
others...

But wait. Many of the sideline personnel have at least some fighting ability. Clarence
Clueful has retired from active fighting because middle age took away his edge, but he still
knows how to use those weapons as well as discuss them. Rodney Rogue’s index card is in the
half-toppled pile of “other people” instead of “fighters” because he isn’t a Council of Light
member, but he does fight. Wouldn’t it be simpler just to merge these, and have one table for all
our human resources?
But then we have fields which are null for almost every entry, such as formal training,
because only rows for full-fledged fighters need them. Those empty fields will lead to a
performance penaltyi which is unacceptable when designing for a secret society charged with
keeping the world safe from the forces of darkness.
So are two tables the right choice? Isn’t that a much worse performance hit, having to search
at least two tables whenever we want to search all monster fighters?
Hmm. Maybe we should put this cowboy database design aside for the moment, to see if we
can learn anything from the normalized approach.

Normalization Forms
We usually speak of five normalization forms. This is a little misleading.ii The standard
numbering is useful for discussion, though, since everyone knows what “Third Normal Form” or
“Normal 3” means without needing an explanation. One might say the names are normalized.
 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p4

First Normal Form

The First Normal Form is really just common sense. It tells us two things: to eliminate
duplicate columns, and that some column or combination of columns should be a unique
identifier.
Eliminating columns which are flat-out duplicates is a no-brainer. Obviously you shouldn’t
create two different columns called “birthdate” and “date of birth,” since a given individual is
generally born only once.
Duplication may be a little less obvious than blatant copies, though. If my monster fighter
table has a firstname field and a lastname field, it might be tempting to have a derived fullname
field to speed up searching. Derived data is allowed at the Normal 1 level.

LAST FIRST FULL NAME DATE OF TRAINING DATE FIGHTING

NAME NAME BIRTH JOINED SKILLS
COUNCIL
Buttkicker Brenda Buttkicker, Brenda 19850123 Knights Templar 20050615 blades, distance
Special Ops weapons,
firearms, martial
arts, munitions,
stakes

But what happens if monster fighter Brenda Buttkicker gets married and takes her husband’s
name? Wouldn’t you, if you were born with a name like “Buttkicker”? Now we need to update
two fields.
We could rely on our application, and all other possible applications ever written by anyone
else, to update both lastname and fullname. We could set up trigger functions to autochange
fullname when lastname is changed, and hope nobody accesses the data through any other
means. Or we could be normal and eliminate the duplication of the lastname and fullname
columns by not having a fullname column, and combine firstname with lastname when we need
a full name.
A “derivation” as simple as bunging two strings together is really just duplication, trying to be
sneaky by wearing a Richard Nixon mask. Duplication, sneaky or otherwise, is disallowed by
Normal 1.
The other requirement of the First Normal Form is that some column or combination of them
be unique in the database. The birthdate field might seem like a good choice, since there are an
awful lot of days in a span of several generations. But what if a pair of twins joins the Council?
Also, as our team grows, chance makes it virtually certain that at least two were born on the
same day. Consider the so-called Birthday Paradox—whenever 23 people are considered, the
odds are better than even at that least two of them share a birthday, so when several hundred
people of one generation are considered the odds of birthdate overlap are unacceptably high.
firstname+lastname+birthdate might be a safe enough choice for uniqueness, at the Normal 1
level.
 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p5

Why all this insistence on uniqueness at the most basic normalization level? Simple. We
need a reliable way to retrieve any given single record.

Second Normal Form

Normal 2 includes all of Normal 1, and goes further. Normal 2 tells us that duplicate data
should be removed and placed in a child table. For instance, I may make a table of known
werewolf clans throughout history, starting with one in Rome.

NAME LOCATION DATE DATE

DISCOVERED ELIMINATED
Dances with Death Rome 12940614 14020819

There’s more than one Rome in the world, so we specify: Rome, Italy, Europe.

NAME LOCATION DATE DATE

DISCOVERED ELIMINATED
Dances with Death Rome, Italy, Europe 12940614 14020819

Let’s add more clans. Here’s one in Athens, Greece, Europe. One in Cairo, Egypt, Africa.
One in London, England, Europe. Hmm, we’re racking up a lot of fields that say, “Europe.”
And here’s one in Rome—heeeey, there was another one in Rome earlier in the list.
Locations can be duplicate data. There’s no sense in having City, Country, and Continent
columns when cities tend to stay put, but we still have to identify which Rome and which Cairo.
We could cram all this into one column, with “Rome, Italy, Europe” as a string.

NAME LOCATION DATE DATE

DISCOVERED ELIMINATED
Dances with Death Rome, Italy, Europe 12940614 14020819
Family of Blood Athens, Greece, Europe 16930321 16980605
Fury with Fur Cairo, Egypt, Africa 19040506 19750302
White Fangs London, United 17890817 18011225
Kingdom, Europe
Zombies Are Lunch Rome, Italy, Europe 19540802 19680812

But sheesh, talk about your duplicate data. We’re making it worse.
 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p6

The smarter approach is to use a location identifier—perhaps modern country codes and the
countries’ postal codes,which are already designed to identify particular areas—and have a
separate locations table which expands “20:11471” into the sixth district of Cairo.

NAME LOCATION DATE DATE

DISCOVERED ELIMINATED
Dances with Death 039:00123 12940614 14020819
Family of Blood 030:17675 16930321 16980605
Fury with Fur 020:11471 19040506 19750302
White Fangs 044:SW8 17890817 18011225
Zombies Are Lunch 039:00172 19540802 19680812

When we use a location ID instead of a human-readable string, we’ve also taken care of the
problem of strings being entered inconsistently. “Philadelphia, United States of America, North
America”? “Philadelphia, PA, U.S.A., North America”? Some other variation? Which is it? It’s
none of those; it’s “001:19131” in a specific part of Philadelphia. By removing duplicate data
into a child table, we’ve also taken care of the problem of sneaky duplicates which don’t actually
look alike. We also distinguish between different districts in large cities, and pre-solve the
problem of what to enter when the Council finds a werewolf clan in London, Ontario, Canada.
Do you feel clever yet? Now we can retrieve data with confidence, knowing that searching
one field for one string will return every werewolf clan which was centered in that area. If we
want to search all of Egypt, we need only search for fields beginning with “020:”. We might
even want to go nuts and prepend a continent number, so that we can trivially search all of
Africa.iii
Normal 2 also encourages us to make use of foreign keys. If I try to enter a new werewolf
clan without a location ID, I should be at least warned and probably prevented. The same is true
if I enter a location ID which isn’t defined in the locations table. If I’m entering a new clan
whose base of operations isn’t yet known, I should use an appropriate reserved ID which is
predefined in the locations table accordingly.
Removing duplicate data led us directly into a better system for categorizing that same data.
Now we’re really getting somewhere.

Third Normal Form

Normal 3 includes all of Normal 2, and goes a big step further. It tells us that a given table
should contain no information which is not directly related to the primary key.
In theory, that sounds like a no-brainer. Of course all the information in a row is related, or
what would it be doing there? But think about the implications. One column is the primary key,
and we should remove any columns which aren’t dependent on the primary key. Every non-key
field must provide a fact about the key.
 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p7

Let’s move on to a table of which weapons work well with which monsters. Our needs
include chainsaws for zombies, bullets for werewolves, and stakes for vampires. Does
composition matter? Absolutely, since vampires require wooden stakes and only a silver bullet
will take out a werewolf. Does the manufacturer matter? Not really; any functional stake or
chainsaw or bullet will do the job. “The stake is wooden” provides a fact we need to know about
stakes when we’re looking up weapons to use. “The stake was made by Oddball Weapons,
LLC,” applies only to some of our stakes, and isn’t relevant to killing the nosferatu currently
battering at our door.

WEAPON MATERIAL MONSTER

axe ANY zombie
bullet silver werewolf
blade ANY zombie
blade metal vampire
blade silver werewolf
chainsaw ANY zombie
cross ANY vampire
garlic ANY vampire
stake wood vampire
sunlight ANY vampire
sunlight ANY werewolf
ANY silver werewolf

We may have other weapons-related tables, such as a current inventory or a list of what
vendors sell what. The manufacturer will be relevant to some of these, but it’s not relevant to the
table of which weapons to use on which monsters, and shouldn’t be stuck in as an extra.
You can use one of your data fields as the primary key, if you know it’s unique. But do you
know it’s unique? Social security numbers are supposed to be unique, but the U.S. government
has issued duplicates by accident. We can still get away with combinations of fields at the
Normal 3 level—although we’re starting to see why we shouldn’t—so perhaps weapon type
combined with weapon composition would work.

Fourth Normal Form

Normal Four includes Normal 3 and adds one more requirement: no multi-value
dependencies. A record shouldn’t contain two or more independent multi-valued facts about an
entity.
Let’s look at that table of fighters again.
 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p8

DATE
FIRST DATE OF
LAST NAME TRAINING JOINED FIGHTING SKILLS
NAME BIRTH
COUNCIL
blades, distance
Knights Templar weapons, firearms,
Buttkicker Brenda 19850123 20050615
Special Ops martial arts,
munitions, stakes
Cutthroat Cathy 19760721 Order of the Stake 20020513 blades, stakes
blades, distance
Killer Karl 19840812 Demon Dojo 20040914 weapons, martial
arts
blades, distance
Martyr Mel 19790304 Demon Dojo 20000912 weapons, martial
arts
Monsterbasher Marvin 19820118 Order of the Stake 20071120 blades, stakes
blades, firearms,
Vampkiller Valerie 19810329 School for Slayers 20030502
stakes
And 50
others...

Consider the Training and Fighting Skills columns. We as humans know that there’s a pretty
close relationship between these two things, but are they guaranteed related? If so, we don’t
need Fighting Skills at all because we can derive it from Training. If not, they’re separate values
(in the logical sense). The Council tells me that they’re separate values, since the Order of the
Stake sometimes drops the ball on training with blade weapons.
So. Let’s pull them out.
First let’s look at a one-to-many relationship. Each fighter attended one and only one training
program, while each training program has many graduates.

UID LAST NAME FIRST NAME TRAINING

Knights Templar
XXX Buttkicker Brenda
Special Ops
XXX Cutthroat Cathy Order of the Stake
XXX Killer Karl Demon Dojo
XXX Martyr Mel Demon Dojo
XXX Monsterbasher Marvin Order of the Stake
XXX Vampkiller Valerie School for Slayers
XXX And 50 others...

I treat the fighter as the key, because the fighter is a “child” of the “parent” training program.
If I treated the training program as the key, I’d either have multiple columns of fighter names for
 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p9

its graduates or I’d still have a separate entry for each fighter-program combination. Why
borrow trouble?
Now we consider the set of fighting skills. This is a many-to-many relationship, so trouble is
borrowing us. To keep the table short, we’ll consider only Brenda and Mel.

LAST
UID FIRST NAME FIGHTING SKILLS
NAME
blades, distance weapons, firearms, martial arts,
XXX Buttkicker Brenda
munitions, stakes
XXX Martyr Mel blades, distance weapons, martial arts

One fighter can have many skills. One skill can be had by many fighters. To avoid multi-
value dependencies, we simply string them out down rows.

UID FIGHTER_ID FIGHTING SKILL

AAA YYY blades
AAA YYY distance weapons
AAA YYY firearms
AAA YYY martial arts
AAA YYY munitions
AAA YYY stakes
AAA ZZZ blades
AAA ZZZ distance weapons
AAA ZZZ martial arts

I can search to find out what skills a given fighter has, or what fighters have a given skill, with
equal facility.
The names, which are now duplicate data, have been extracted into a child table. Each fighter
has an assigned ID number. I’ll also apply these FIGHTER_ID numbers in the table of training
programs, table of biographical data, and whatever else we need. I can use the same ID number
when looking up Cathy Cutthroat’s fighting skills for mission assignment, her training program
to find her fellow alums, or her birthday to plan her surprise party.
I’ve left the fighting skills as text for readability, but the duplication of data is pretty obvious.
Fighting skills will probably also have ID numbers, unless we can guarantee that all “blades” are
created equal and that there’s nothing else we’ll ever need to record about those.
Keen observers will note that we now require a guaranteed-unique ID in a single field. This
isn’t always mission-critical. In SQL, you can simulate it with views. Really, though, why not
have a guaranteed-unique ID?
 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p10

We now have a required primary key in a single field, with no more of this namby-pamby
nonsense about some combination of fields being unique. A good practice is to have a column
just to hold a unique numeric key. Another good practice is to let the database assign a number,
via SQL’s newid() function.
If you want IDs which are universally unique and not just unique to your table or database,
another choice is to follow the UUID standard as defined in RFC 4122. Your programming
language of choice probably already has a library/class/module/whatever to generate them (e.g.,
java.util.UUID).
Some databases, such as Postgres, support UUID as a native datatype. For others, just
remember that it’s 128 bits by definition and go from there. Some people simply store it as a
CHAR(36), while others get fancier (e.g., stripping hyphens and performing hex conversions) to
store in other formats.
A Universal UID isn’t necessary to meet Normal 4, though. The UID need only be unique to
the table, so if duplicates in other tables aren’t a problem in your design, letting the database
assign the ID is an admirably lazy technique.
Marvin+Monsterbasher+19820118 looks likely to stay unique, but I’m not convinced that
stake+wood is good enough. I’m taking all of the Council’s tables to Normal 4.
Let’s look briefly at both approaches to a unique ID. If I choose the UUID approach, my
primary keys will now all be something like
12345678-1234- 5678-1234-567812345678
and my other fields tell me some fact about that—starting with what the heck it is. As you
can see, it’s a bit much for a human to process visually or mentally. It need never be displayed in
reports intended for human eyes, though.
If I let the database assign non-universal UIDs, the numbers will be smaller—probably small
enough for humans to manage, so that people could memorize their own monster fighter ID
numbers. If I want the UIDs human-usable and don’t need them to be unique in the entire
universe, that’s probably the right approach. If any Council members complain about feeling
like numbers instead of people, I’ll just mumble something about the computer requiring it.

Fifth Normal Form

Normal 5 doesn’t come up often. At least in theory, any data set normed to Normal 4 usually
conforms to Normal 5 without further work. There are those times, though, and the Council has
presented me with some things I should consider.
There is no such thing as a wooden chainsaw, for obvious reasons. I can let my application
carry the burden of preventing users from accidentally entering a wooden chainsaw as a new
weapon type, or I can bite the bullet and tackle Normal 5. Also, in our monster fighters table, we
have a near-opposite situation—any fighter who has finished one of the several anti-monster
training programs is presumed competent in the particular fighting techniques taught by that
program. I can let my application fill in duplicate fields of techniques, or I can normal up and
conform to Normal 5.
 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p11

A data set conforms with the Fifth Normal Form if and only if every join dependency in it is
implied by the candidate keys.
Um, what?
When real-world constraints mean that some combination of allowed values is not a valid
combination, that combination must be disallowed by the logical data structure itself in order to
meet Normal 5. For instance, a real-world situation may be such that W=X means that Y can
never equal Z, but the existing Normal 4 structure of the table doesn’t disallow it because each
field entry is valid. We could apply a constraint, but that’s cheating. We’re talking about the
logical data design itself.
On the other side of the coin, a real-world constraint may mean that value A with value B
always implies value C. The Normal 4 structure would duplicate value C for each such entry,
even though it can be derived from the existence of values A and B. This isn’t derived data in the
usual sense—it’s not A times B, for example. It’s inferable because something in the real world
tells us that it must always be that way.
To meet Normal 5, we must set up the tables such that Y=Z can never occur in the logical data
design, and we must remove C on the grounds that it can be inferred from A and B.
Um. What?
It’s hard to understand this abstractly, so let’s talk about the Council of Light’s weapon data.
This is going to get detailed, but we don’t have to consider every possible combination. That’s
the Council’s job. Our job is to create a data model which accurately represents the data they’ll
put in.
Let’s start with the wooden chainsaw problem. A monster is vulnerable to certain weapon-
material combinations. We may choose to model this data by saying that it is vulnerable to some
weapons and some materials. We derive its weapon-material vulnerabilities by checking its
weapon vulnerabilities and its material vulnerabilities. Let’s look at our weapons table again:
 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p12

UID WEAPON MATERIAL MONSTER

XXX axe ANY zombie
XXX bullet silver werewolf
XXX blade ANY zombie
XXX blade metal vampire
XXX blade silver werewolf
XXX chainsaw ANY zombie
XXX cross ANY vampire
XXX garlic ANY vampire
XXX stake wood vampire
XXX sunlight ANY vampire
XXX sunlight ANY werewolf
XXX ANY silver werewolf

We’re seeing a fair bit of duplication. Consider the werewolf entries—all but one are for
silver things. We have an ANY silver entry, but bullets are there separately because they’re the
weapon of choice, and we also have “blade” for consistency when searching for monsters who
can be hurt by blades. This isn’t needful duplication. Given that monster is werewolf and
material is silver, we can derive the fact that the weapon will do damage in its usual manner.
Normal 5 tells us that we should remove any data which can be derived from other data—
such as “silver [anything] hurts werewolves.”
Is this universal? Is it something that will always be true? If so, this means that I split my
weapons table into three tables. One is for weapon type and weapon composition, one is for
weapon type and monster vulnerability, and one is for weapon composition and monster
vulnerability.

UID WEAPON MONSTER UID MATERIAL MONSTER UID WEAPON MATERIAL

XXX axe zombie XXX silver werewolf XXX axe metal
XXX stake vampire XXX wood vampire XXX axe silver
XXX stake metal
XXX stake wood

One tells me that vampires are vulnerable to stakes. Another tells me that vampires are
vulnerable to wood. The third tells me that stakes of wood are available. From the first two, we
derive the fact that wooden stakes are handy anti-vampire weapons. Meanwhile,
“wood+chainsaw” is coming up blank from the “weapon type and weapon composition” table.
What if someone creates a chainsaw with wood-tipped chain links specifically for anti-
vampire use? That’s pretty cool, and we have the case pre-covered. If we create an entry for a
 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p13

chainsaw whose relevant material is wood, it’ll automatically come up as a possible anti-vampire
weapon because of the wood vulnerability.
This leads to some problems, though. Vampires are vulnerable to the material ANY—but
only if it’s in the shape of a cross. What does that mean for our material-monster table? There
aren’t enough real-life material-monster constraints to make this a good data model, and so our
previous table doesn’t need to be split further to meet Normal 5—at least, not in that particular
way. Still, humor me while I carry this a bit further.
Our hypothetical weapon-material table also can’t represent reality as well as we might wish.
Wooden sunlight, anyone? Are there any possible composition values for sunlight other than,
well, “sunlight”? What’s the weapon and what’s the material?
But sunlight, although deadly to vampires, isn’t really a weapon as such. You can’t carry it
around and deploy it. Perhaps it shouldn’t be in our weapons table. We can create a separate
table for environments, including those which are so hostile to the monster that they actively
harm it.
Nota bene: Applying Normal 5 breaks data up into smaller and smaller tables. Splitting up
our data into such small and specialized tables forces us to think about edge and corner cases.
Hmm. That’s good to know. Our first attempt at further breakdown didn’t work out so well,
but let’s follow this thought process further. What won’t fit? Why won’t it fit?
A moment ago we were thinking about crosses and vampires. From the most elaborate silver
crucifix to a couple of sticks properly tied, any cross can hold a vampire at bay. Composition
“ANY” is needed to represent that.
With werewolves, on the other hand, only the silver crucifix would be useful and then only if
we can bash the monster on the noggin with it. In the werewolf case, it’s the silver composition
that matters and we’ll need the semi-generic weapon “blunt instrument.” The data model’s
failure taught us something about what we’ll need for a successful one.
We could also think about axes. Only cold iron and cold steel work against evil fey, but that’s
a metal vulnerability. Anything sharp enough to cut flesh can dismember a zombie. A silver axe
would be too soft to cut well but would still hurt a werewolf. Vampires can be cut by anything
sharp but must be beheaded to kill them that way. A given axe could fall into any or all or none
of the appropriate categories.
 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p14

UID WEAPON MATERIAL MONSTER

XXX axe ANY zombie
XXX axe gold
XXX axe iron fey, zombie
XXX axe metal zombie
XXX axe silver werewolf,
zombie
XXX axe steel fey, zombie
XXX axe wood

Hmm. Metal and iron aren’t at the same level; most of those things are metals. And we
aren’t distinguishing between any old iron and the cold-worked iron needed to hurt the fey
creatures. Perhaps what we really need is a hierarchy:

Everything -> inorganic -> metal -> iron -> cold

 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p15

UUID KINGDOM CONTAINS

XXXX ANY
XXXX animal YYYY,
YYYY,
YYYY
XXXX energy YYYY,
YYYY UUID INORGANIC CONTAINS

XXXX YYYY, YYYY ANY

inorganic
YYYY, ZZZZ, ZZZZ,
YYYY, YYYY crystal ZZZZ, ZZZZ,
YYYY, ZZZZ
YYYY,
ZZZZ, ZZZZ,
YYYY
ZZZZ,
XXXX vegetable YYYY,
YYYY metal ZZZZ,
UUID METAL TYPES
YYYY, ZZZZ ZZZZ ANY
YYYY
ZZZZ, ZZZZ, ZZZZ copper
YYYY other ZZZZ, ZZZZ,
ZZZZ gold
ZZZZ
AAAA,
ZZZZ iron BBBB
CCCC,
ZZZZ silver DDDD
AAAA,
ZZZZ steel BBBB

Now we’re really getting somewhere. Mineral refers us to metal, which refers us to silver,
which can refer us to blessed or unblessed if necessary. Note that I’ve traded in my table-unique
IDs for universally unique IDs. The big win is that “energy, sunlight” and “mineral, metal,
silver” can each be listed as a monster vulnerability, even though they’re on different hierarchical
levels.
For we designing the database, the task isn’t to think up every possible combination. That’s
the Council’s job. Our job is to think of what kinds of things we need to record, and examples
which push the boundaries of standard rules help us do that.
You’ve probably noticed that we have a one-to-many or many-to-many relationship a single
row, listing multiple children in a single field. I left it that way for readability, but a truly
normalized table would consist of nothing but IDs until the leaf level(s). A slice might look like
this:

UUID PARENT_ID CHILD_ID

XXXX AAAA QQQQ
 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p16

XXXX AAAA RRRR

XXXX BBBB QQQQ

Woo. We’re rapidly moving further away from the human-readable. We may not choose to
do it precisely this way.
Whether or not we choose to implement Normal 5, thinking it through is leading us to some
good realizations. A team of vampire fighters, trapped in a cellar at midnight, would not be
amused to learn that sunlight is a useful option. A team of werewolf hunters, on the other hand,
may be pleasantly surprised to realize that the silver part of their vampire-hunting kit can be
repurposed.
I don’t recommend Normal 5 for most real-world situations even when it applies smoothly,
since it doesn’t allow for exceptions. Those real-world constraints on which we relied may turn
out not to constrain universally and forever. What if an item is made of silver but is so small and
light that it can’t conceivably hurt a werewolf?
“Anything silver hurts werewolves” and “all crosses hurt vampires” may be true in theory, but
the real world seldom matches theoretical constraints. A thrown dime isn’t a useful anti-
werewolf missile. A cross made by crossing your forefingers will only slow a vampire down if
he’s laughing too hard to attack you. A database for real use must let us store data for real cases,
including exceptions to general rules.
Still, even without implementing it, trying to apply the Normal 5 process has shown us that
improbable combinations should issue warnings, that some things we think of as weapons should
be treated differently from others, that materials of composition are a hierarchy rather than a
simple list, and that weaponry will require a great deal more attention than we realized back in
Normal 4.
By using Normal 5 as a thought exercise, we get many of its planning benefits without its
drawbacks of numerous microtables and performance slowdowns.

Sixth Normal Form

Hey, didn’t she say “five” earlier? Yeah, but I’m going to mention this for completeness.
You may have heard of a Sixth Normal Form. This term is used in two different ways, and
both of them are for academic use. It’s intended to break down data relations until they’ve been
reduced to irreducible components. If you’re designing real databases for real use, ignore it.
Both of it. Please.

Denormalization
Denormalization doesn’t mean not normalizing. It’s a step in the process. First we normalize,
then we realize that we now have hundreds or thousands of small tables and that the performance
 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p17

cost of all those joins will be prohibitive, and then we carefully apply some denormalization
techniques so that our application will return results before the werewolves have overrun the city.

Deciding How Far to Normalize

I believe that running all the way up to Normal 5 can be a useful thought exercise for any
database which is being designed instead of just tossed together. But when it’s time to do the
grunt work of mapping all the actual tables and columns, we first decide how far we need to go.
Even a spreadsheet should comply with Normal 1. There’s really no excuse for duplicate
columns and non-unique entries. If your hardware can’t handle that and your data is
understandable without it, do you need a real database at all? Searching a directory of text files
might serve you better.
Normal 2 is a good idea for all but the smallest and most casual of databases. If you
coordinate a group of half a dozen fighters who only battle zombies and operate entirely in
Boise, then duplicate data may not be a problem and you may be able to ignore Normal 2.
Everyone else should conform to it.
Normal 3 is necessary, at least in the long term, for any business operation or any other
operation where the data is critical. It may seem a bit hardcore, but it’s vital for any organization
which has more data than it could reasonably process by hand. That’s the crossover point at
which we must treat the database as a vital part of the organization’s mission.
Normal 4 isn’t always necessary, but if you’ve done Normal 3, why not have a guaranteed-
unique ID for each record? Unlike some of the other steps, this one is actually a performance
win. It’s faster to look up one field than to check several to see if the combination’s unique, and
you still have the option of searching by other means. Stringing data down rows instead of
across columns may lead to performance slowdown, but there are ways to address this, and we’ll
look at some of them. Normal 4 is an excellent idea for any organization with large amounts of
data—again, because keeping the data straight is vital to the organization’s mission.
Normal 5 is rarely necessary. It reaches into rarified heights where the purity of the data
model is more important than the real-world usability. Except in rare cases where permanent and
inflexible real-world constraints match the logical data design beautifully, you’ll probably end up
needing to denormalize anyway. Think it through, because you’ll realize some good stuff, but
don’t worry about meeting it to the letter. If you did Normal 4 correctly, you probably already
met it.
To summarize— at least Normal 2 for tiny organizations, and at least Normal 4 for any
operation where the data’s critical.
I’ll be conforming to Normal 4 for the Council of Light’s final database design, although
thinking about Normal 5 did lead me to some realizations I’ll incorporate into my design.

Should I Denormalize?
Since normalization is about reducing redundancy, denormalizing is about deliberately adding
redundancy to improve performance. Before beginning, consider whether or not it’s necessary.
 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p18

● Is the system’s performance unacceptable with fully normalized data? Mock up a

client and do some testing.
● If the performance is unacceptable, will denormalizing make it acceptable? Figure out
where your bottlenecks are.
● If you denormalize to clear those bottlenecks, will the system and its data still be
reliable?
Unless the answers to all three are “yes,” denormalization should be avoided.
Unfortunately for me, the Council of Light has some pretty old hardware and can’t or won’t
upgrade to newer and more efficient SQL software. I’d better get to work.

Denormalization Strategies

Materialized Views
If we’re lucky, we won’t need to denormalize our logical data design at all. We can let the
database management system store additional information on disk, and it’s the responsibility of
the DBMS software to keep the redundant information consistent.
Oracle does this with materialized views, which are pretty much what they sound like—SQL
views, but made material on the disk. Like regular views, they change when the underlying data
does. Microsoft SQL has something called indexed views, which I gather are similar although
I’ve never used them. Other SQL databases have standard hacks to simulate materialized views;
a little Googling will tell you whether or not yours does.

Database Constraints
The more common approach is to denormalize some carefully chosen parts of the database
design, adding redundancies to speed performance. Danger lies ahead. It is now the database
designer’s responsibility to create database constraints specifying how redundant pieces of data
will be kept consistent.
These constraints introduce a tradeoff. They do speed up reads, just as we wish, but they slow
down writes. If the database’s regular usage is write-heavy, such as the Council of Light’s
message forum, then this may actually be worse than the non-denormalized version.

Double Your Database Fun

If enough storage space is available, we can keep one fully-normalized master database, and
another denormalized database for querying. No app ever writes directly to the denormalized
one; it’s read-only from the app’s point of view. Similarly, no app ever reads directly from the
normalized master; it’s write-only from the app’s point of view.
This is a good, if expensive, plan. The fully normalized master database updates the querying
database upon change, or the querying database repopulates itself on a schedule. We no longer
have slower writes, and we still have faster reads. Our problem now becomes keeping the two
 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p19

database models perfectly synchronized. That itself takes time, of course, but your particular
situation may be such that you can do this.
Most commonly, the denormalized version populates itself by querying the normalized master
on a controlled schedule. It will thus be the only thing which ever reads from the normalized
master.
The danger is that data returned from the denormalized query base may be inaccurate if the
data has changed in the normalized master but the denormalized query database hasn’t
repopulated itself yet. If your data doesn’t generally change in a heartbeat, this may be
acceptable. If the future of all humankind hinges on the currency of your data, this approach
may not be acceptable.
Another possibility is for the normalized master to change the denormalized query database
whenever the normalized master is changed. This requires a complex set of triggers interfacing
the normalized and denormalized data designs, and the trigger set must be updated when the
denormalized design is altered. The need to handle triggers will slow down writes, so this isn’t a
good option for write-heavy applications, but the upside is that your two databases are much less
likely to be out of synch.
Space requirements often prohibit this strategy entirely.

Let the App Handle It

We add redundancies to our database, but let the application worry about synchronizing
redundant fields. We avoid the performance hit on writes, but now we’re relying on a different
piece of software to do the right thing at all times.
If this is the choice, document it thoroughly and unavoidably. Make a stern warning and a list
of the redundancies the first thing any future developer will see. It may also help if you include a
threat about noncompliant apps contributing to a future in which the world is overrun by
creatures of the night.

Denormalization Tactics
All right, so those are the basic approaches. All of them have one thing in common, which is
that they deliberately introduce redundancies. How do we decide what redundancies to
introduce?
Simple. We find out exactly what is causing those performance problems we noted—you
know, the ones which must be present in order for us to consider denormalization. Then we
introduce redundancies which will make those particular causes go away.
The two most common bottlenecks are the cost of joining multiple tables and, on very active
systems, the activity on a particular table being too high.

Pre-Joined Tables
Is it the joins? Which ones?
 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p20

Perhaps we find that our users do a lot of queries which require joining the vampire table and
the zombie table. If so, we could combine them into a single pre-joined table of the undead.
If our users often search all available personnel, joining combatants and former combatants
and noncombatants, then we may want to revert to our idea of keeping all Council members and
other helpers in a single personnel table.
This table is essentially the materialization of a join, so we populate if from the original tables
either on a schedule or via triggers when changing the originals.

Mirror Tables
If a particular table is too active, with constant querying and updating leading to slowdowns
and timeouts and deadlocks, consider a miniature and internal version of “Double your database
fun.” You can have a background and foreground copy of the same table. The background table
takes only writes, the foreground table is for reading, and synchronizing the two is a low-priority
process that happens when less is going on.
The obvious problem is that the data you’re reading has no guaranteed currency. In fact,
given that we’re doing this because a table is particularly active, it’s fairly likely not to be
current. It’s something to consider when the data in question doesn’t require up-to-the-second
currency, though.

Report Tables
Each day, between sunset and midnight, the Council’s system is pummeled with queries on
how to take out particular monsters. There’s a lot of joining of weaponry tables, and many users
are making the same queries to the same tables.
I can address this by making redundant tables just for these extremely common reports. My
table on how to take out a vampire will be modeled after the desired report rather than around the
logical data design, and might look something like this:

UID WEAPON HOW TO USE EFFECT

XXX cross expose monster to it monster retreats in pain
XXX stake, wood stab monster through the heart kills monster
XXX sunlight expose monster to it kills monster
XXX water, fresh, running put between self and monster barrier to monster

Except for the UID for internal use, this table contains exactly what the user wanted to know
and is therefore extremely fast to query.
The report table doesn’t take writes from apps. It updates itself from the logical data design
tables, regularly during off-peak hours and during peak hours as system resources permit.
 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p21

Split Tables
We may have distinct groups of users who query tables in distinct ways. Remember our table
of werewolf clans? Monster fighters and coordinators may be interested only in clans which
haven’t been exterminated, while historians studying the effectiveness of extermination methods
may be interested only in those which have. I could split that single table in two. The original
table could also be maintained as a master, in which case the split tables are special cases of
mirror tables (above) which happen to contain a subset instead of a complete copy. If few people
want to query the original table, I could maintain it as a view which joins the split tables and treat
the split tables as masters.

UID NAME LOCATION DATE DATE

DISCOVERED ELIMINATED
XXX Dances 039:00123 12940614 14020819
with
Death
XXX Family 030:17675 16930321 16980605
of Blood

UID NAME LOCATION DATE

DISCOVERED
XXX Fangs 01:45701 20091001
and Fur
XXX Howl 044:SW1 20100105

That’s a horizontal split, pulling rows out into different tables. For other needs we might do a
vertical split, keeping all the same rows/keys but with each table having separate sets of columns
of information.
This is only worthwhile when there are distinct kinds of queries which are, in effect, already
treating the table as more than one table.

Combined Tables
Instead of splitting one table into several, we might combine several into one. If tables have a
one-to-one relationship, it might speed performance to combine them into one table even if that
isn’t part of the logical database design. We might even combine tables with a one-to-many
 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p22

relationship, although that would significantly complicate our update process and would likely
work out to be a net loss.
For instance, I might combine the table of monster fighters with the table of training programs
on the grounds that there are duplicate columns (school names, skills taught/learned). That could
be useful for high-level coordination, which might require tracking current munitions training
programs by evaluating both munitions programs and the fighters rated as skilled in munitions.
A join would serve, but we might go with a combined table if we find that we’re getting many
such queries.
Combined tables differ from joined tables (above) in that they already share some data in
some relationship, while any two tables which are frequently joined are potential candidates for
joined tables. They also differ in that a joined table usually populates itself from the co-existing
normalized tables, while a combined table will usually be the master copy.
So what’s the difference between splitting tables and never combining them in the first place?
Or between combining tables and just not splitting them? Hey, I never said this was a science.
It’s more of an art. Or possibly a craft. Whatever it is, it’s definitely not for sissies. The
redundancies you choose to introduce must be based on the particular bottlenecks in your own
specific data set.

Index Tables
These are nice for searching hierarchies. Remember how Normal 5 taught us that weapon
composition is actually a hierarchy of materials rather than a simple list? Well, now we’re stuck
with searching that hierarchy by searching all the leaves under particular parent nodes and
combining with a bunch of union statements.
Or are we? We could make a metatable consisting of nothing but parent-child pairs and any
data which would speed searching. Like maybe this:

UUID PARENT_ID CHILD_ID LEVEL LEAF

XXXX AAAA BBBB 1 N
XXXX AAAA CCCC 1 N
XXXX BBBB QQQQ 2 Y

This table gets large fast, but it’s all numeric and Boolean, so searching is a breeze. It’s
simply an index for that tree of materials we made. Technically it’s derivable data (below), but
this is a special case which is almost always worthwhile when more than a minimal hierarchy
must be searched.
It can repopulate itself on a schedule, or changes to the main databases can trigger changes in
the index table.
 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p23

Repeating Groups
Let’s say that my table of monster fighters’ skill eventually shook out like this:

UID FIGHTER_ID SKILL_ID RATING

XXX AAA 10 5
XXX AAA 15 3
XXX BBB 12 4
XXX BBB 23 3
XXX BBB 17 4

This can hold an infinite number of skills. I could instead choose to compress rows, stringing
information across columns instead of down rows:

UID FIGHTER_ID SKILL_ID_1 RATING_1 SKILL_ID_2 RATING_2 SKILL_ID_3 RATING_3

XXX AAA 10 5 15 3
XXX BBB 12 4 23 3 17 4

Is this a good idea?

Well, it reduced the number of rows to search. The one-row-per-fighter structure seems like it
would be efficient, since I won’t normally need to do operations within a single row. The data in
the new row format is often accessed collectively. Those are all plusses.
On the minus side, I’ve artificially limited the number of skills a fighter can have. There’s no
stable and predictable number of fighting skills, so putting a limit on it is a big minus. Even if I
make the number unreachably huge, now I’ve got a table with tons of nulls.
This could be a good idea for data where there’s an actual hard limit on the number of fields,
though. If I’m making a report table (above) for monster activity within the past 12 months, a
limit of 12 on the months is quite reasonable.

Derivable Data
In rare cases, it’s the cost of doing a calculation which is prohibitive. These are usually cases
in which the algorithm for calculation is enormously complex and/or is itself dependent on
expensive queries. In such cases, it might be worth storing value C even though value A with
value B always works out to value C.
Changes to A or to B can trigger a function to rederive C. If C needs to be retrieved often, this
can be a lot cheaper than rederiving it every time it’s needed.
 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p24

Index tables (above) are derivable data, but they’re different in function. Those are for
internal use only, containing no data that anyone is likely to want to use directly. These are
derivable data which might be presented to a user.

Redundant Data
Yes, plain old duplicated data. A violation of Normal 1.
If we get a large number of queries which take most of their data from one table but have to
join with a single column from another table, it may be worth duplicating that one column in the
first table. This is almost never a good idea, since it goes against the basic principal of
normalization and brings back all the problems we were trying to avoid in the first place. It may
be worth a shot if normal database access is downright unbearable and some other constraint
makes a join table (above) unacceptable.

Denormalization in Hyperspace:
Star Schemas, OLAP Cubes, and Other Weirdness
These are elegant, but a great deal of design work because they require math that most of us
never learned. Modeling data into cubes? It sure sounds interesting, but it also sounds
suspiciously like one of those approaches where the approach turns into the real project.
Lacking a Ph.D. in mathematics, I’m as likely to make things worse as I am to make them
better. These things are fun to read about, but not suitable for most real-world deployments. If
you think you’re up for the task, Wikipedia is either a great starting place or a much-needed
reality check.

Conclusion
How does this help me with my original problem? I started with one question: Should all
Council members be in one table, or should fighters and noncombatants be treated separately?
Well, I’d just about concluded that I should have several separate tables for people in different
roles and use a Postgres hack to simulate a materialized view for searching all personnel at once.
That will avoid both the join performance hit and the many-nulls performance hit. After going
through the normalization and denormalization process, I see that that’s the best answer.
But then I realized that the Council of Light never actually agreed to pay me. Tightwads. If
they can’t cough up a few bucks to save the world more efficiently, they can keep poring over
their dusty tomes forever, and I don’t care if all Hell breaks loose.
Um. Perhaps I should rephrase that . . . .
 Data Normalization, Denormalization, and the Forces of Darkness / Hollingsworth / p25

Endnotes
i SQL nulls are just plain unintuitive. SQL has an odd aggregation of behaviors with respect to its nulls. It’s never safe to
extrapolate from what you know about them to what they ought to do in other cases. Other white papers have been
written entirely on SQL’s nulls.

Here are a couple of nutshell reasons why using lots of null values is not a big performance win: A null takes up the
entire width of the column, so that storing the string “NULL” would actually be faster in many cases. If you use an IS
NOT NULL condition, your SQL performance will be normal, but if your search allows for null values—as in our search
for all monster fighters, regardless of training or active duty status—then performance slows measurably.

At least one major SQL vendor recommends never using the standard-mandated null, and always fudging via some other
reserved value.

ii There’s a “three and a half,” better called the Boyce-Cobb standard, which is a little more restrictive than the Third
Normal Form. Anything which complies with the Fourth Normal Form usually conforms to the Fifth Normal Form
automatically, without further design effort, leading some to wonder if they really deserve different whole integers.
There are also two different Sixth Normal Forms. We’ll ignore the weirdest stuff and stick to the standard five; you can
always do your own further research if you’re really interested.

iii I’m using this colon-delineated format because country codes and postal codes already exist and are standardized. I see
no sense in reinventing the wheel. As long as I’m keeping them in one field, a non-numeric delineator works best. This
system is guaranteed unique in theory, since each country has a separate country code and each country bears
responsibility for keeping its internal postal codes unique. I’d have to reserve values for “location unknown” and for
remote locations without postal codes.